Hybrid DE Algorithm With Adaptive Crossover
Operator For Solving Real-World Numerical
Optimization Problems
Gilberto Reynoso-Meza, Javier Sanchis, Xavier Blasco and Juan M. Herrero
Instituto Universitario de Automática e Informática Industrial
Universitat Politècnica de València
Camino de Vera s/n 46022, Valencia, Spain
Abstract—In this paper, the results for the CEC 2011 Competi-
tion on testing evolutionary algorithms on real world optimization
problems using a hybrid differential evolution algorithm are
presented. The proposal uses a local search routine to improve
convergence and an adaptive crossover operator. According to
the obtained results, this algorithm shows to be able to find
competitive solutions with reported results.
Index Terms—Differential Evolution algorithm, parameter se-
lection, CEC competition.
I. I NTRODUCTION
In this paper, a hybrid differential evolution algorithm with
a local search (LS) technique and an adaptive mechanism for
crossover operator has been used to solve a set of single-
objective optimization problems. Such problem set was defined
for the CEC-2011 Competition and Special Session on testing
evolutionary algorithms on real world optimization problems.
The problem definitions and evaluation criteria are defined
in [1]. This competition includes a set of 13 problems,
with several diverse topics, which includes control systems,
engineering design and aerospace applications (see table I).
The problem set consists of several single objective opti-
mization statements, with different characteristics, number of
decision variables and constraints. In general, a constrained
single objective optimization is stated as follows:
min 𝐽(𝒙) ∈ ℜ1
𝒙∈ℜ𝑛
𝑠.𝑡.
𝑔𝑖 (𝒙) ≤ 0, 𝑖 = 1, . . . , 𝑝
ℎ𝑗 (𝒙) = 0, 𝑗 = 𝑝 + 1, . . . , 𝑟. (1)
𝑛 popular mechanisms is the binary crossover. For each target
where 𝒙 ∈ ℜ is known as the decision vector of 𝑛
variables. A common approach to constraint handling in
optimization problems is by means of penalty functions. An
aggregate cost function is built using the original cost function
and a penalty value according to the amount of the violation
constraint. With this in mind, some of the problems have been
re-formulated in the form:
min 𝔍(𝒙) = 𝐽(𝒙) + 𝑓𝑝𝑒𝑛𝑎𝑙𝑡𝑦 (𝒙, 𝑔𝑖 (𝒙), ℎ𝑗 (𝒙)) ∈ ℜ1
𝒙∈ℜ𝑛
(2)
978-1-4244-7835-4/11/$26.00 ©2011 IEEE 1551
Several problems in the set have been subject of research
in order to generate new evolutionary algorithms strategies,
due to their complexity [1]. Sometimes, such algorithms are
highly specialized in a particular kind of problem, and the
search for an algorithm of general purpose is worthwhile. Due
to this fact, in this paper a proposal based on Differential
Evolution (DE) algorithm is presented. The remainder of this
paper is as follows: in section II the algorithm is presented.
The experimental setup used to solve the proposed problems
is shown in section III. In section IV the obtained results are
commented and finally, some concluding remarks are given.
II. E VOLUTIONARY OPTIMIZATION ALGORITHM
Nowadays, DE algorithm [2]–[4] is perhaps one of the most
powerful evolutionary algorithms to solve real parameter op-
timization problems [4]. Several state-of-the-art evolutionary
algorithms use DE as evolutionary mechanism [5]–[12]. For
this reason, DE algorithm is used as proposal in this work.
A. Differential Evolution Review
Generally speaking, DE algorithm has three main steps:
1) Mutation: At generation 𝑘 for each target (parent) vector
𝒙𝑖 ∣𝑘 , a mutant vector 𝒗 𝑖 ∣𝑘 is generated according to equation
(3):
𝒗 𝑖 ∣𝑘 = 𝒙𝑟1 ∣𝑘 + 𝐹 (𝒙𝑟2 ∣𝑘 − 𝒙𝑟3 ∣𝑘 ) (3)
Where 𝑟1 ∕= 𝑟2 ∕= 𝑟3 ∕= 𝑖 and 𝐹 is known as the Scaling
Factor.
2) Crossover: This operator is crucial due to its sensitivity
to face separable and non-separable problems. One of the most
vector 𝒙𝑖 ∣𝑘 and its mutant vector 𝒗 𝑖 ∣𝑘 , a trial (child) vector
𝒖𝑖 ∣𝑘 = [𝑢𝑖1 ∣𝑘 , 𝑢𝑖2 ∣𝑘 , . . . , 𝑢𝑖𝑛 ∣𝑘 ] is calculated as follows:
{ 𝑖
𝑣𝑗 ∣𝑘 𝑖𝑓 𝑟𝑎𝑛𝑑(0, 1) ≤ 𝐶𝑟
𝑢𝑖𝑗 ∣𝑘 = (4)
𝑥𝑖𝑗 ∣𝑘 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
where 𝑗 ∈ 1, 2, 3 . . . 𝑛 and 𝐶𝑟 is the Crossover probability
rate. In its standard version, at least one dimension must be
exchanged. For separable problems, low values of 𝐶𝑟 are
recommended while for non-separable problems, high values
are suggested [3].
 TABLE I: Problem set definitions. {
Problem Variables Comments 𝑖
𝒖 𝑖 ∣𝑘 𝑖𝑓 𝑓 (𝒖𝑖 ∣𝑘 ) ≤ 𝑓 (𝒙𝑖 ∣𝑘 )
𝒙 ∣𝑘+1 = (6)
P01 6 Parameter estimation for 𝒙𝑖 ∣𝑘 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
frequency-modulated (FM) sound
waves.
B. DE variant proposal
P02 30 Lennard-Jones Potential Problem.
P03 1 Bifunctional catalyst blend optimal The proposal in this work is to use a hybrid DE algorithm
control problem. with local search techniques and an adaptive parameter value
P04 1 Optimal control of a non-linear
stirred tank reactor. for 𝐶𝑟. This proposal will hereafter be denoted as 𝐷𝐸 − Λ𝐶𝑟 .
P05 30 Tersoff potential function mini- This adaptive mechanism combines the binary crossover and
mization problem (instance 1). the line recombination, to face the diversity in the problem set.
P06 30 Tersoff potential function mini-
mization problem (instance 2). Also, a population refreshment mechanism is used to avoid
P07 20 Spread spectrum radar polly phase stagnation.
code design. 1) Local search strategy: A sequential quadratic program-
P08 7 Transmission network expansion
planning problem. ming (SQP) routine is incorporated to improve convergence
P09 122 Large scale transmission pricing as a local search mechanism. The SQP routine is executed in
problem. every child of the population with a probability 𝛼𝐿𝑆 . Local
P10 12 Circular antenna array design prob-
lem.
search strategies has shown to improve the performance of the
P11.1 120 Dynamic economic dispatch prob- classical DE algorithm [14]–[16].
lem, instance 1 2) Adaptive Parameter values: Triangular distributions Λ𝐹 ,
P11.2 216 Dynamic economic dispatch prob- Λ𝐶𝑟 are used for both values 𝐹 and 𝐶𝑟 to calculate 𝑓𝐹 (𝒙)∣𝑘
lem, instance 2
P11.3 6 Static economic load dispatch and 𝑓𝐶𝑟 (𝒙)∣𝑘 for each child in the evolutionary process at
problem, instance 1 generation 𝑘. In the case of 𝐹 , it is used to incorporate some
P11.4 13 Static economic load dispatch dither in the algorithm. Whilst the triangular distribution Λ𝐹
problem, instance 2
P11.5 15 Static economic load dispatch is fixed, the triangular distribution Λ𝐶𝑟 is adapted after a
problem, instance 3 number 𝛾𝑠𝑢𝑐𝑐𝑒𝑠𝑠 of recorded success. It will be said to have a
P11.6 40 Static economic load dispatch success if a child substitutes its parent in the next generation.
problem, instance 4
P11.7 140 Static economic load dispatch
The minimum, maximum and medium value on such set of
problem, instance 5 success is used for this purpose. The ideas behind this adaptive
P11.8 96 Hydrothermal scheduling problem, mechanism for the crossover are:
instance 1
P11.9 96 Hydrothermal scheduling problem, ∙ Be able to detect a separable problem, choosing a bino-
instance 2 mial crossover operator with low values for 𝐶𝑟.
P11.10 96 Hydrothermal scheduling problem, ∙ Be able to detect non-separable problems, choosing a
instance 3
P12 26 Spacecraft trajectory optimization binomial crossover operator with high values for 𝐶𝑟.
problem. ∙ Be able to detect a strong-dependency on decision vari-
P13 22 Spacecraft trajectory optimization ables, and use a non-rotationally invariant line recombi-
problem.
nation.

The binomial crossover (for 𝐶𝑟 < 1) is not a rotationally

invariant operator. For functions with a high dependency
among search variables, DE could face some difficulties [4],
and a rotationally invariant line recombination operator [13]
was suggested as an alternative.
For each target vector 𝒙𝑖 ∣𝑘 and its mutant vector 𝒗 𝑖 ∣𝑘 , a
trial (child) vector 𝒖𝑖 ∣𝑘 = [𝑢𝑖1 ∣𝑘 , 𝑢𝑖2 ∣𝑘 , . . . , 𝑢𝑖𝑛 ∣𝑘 ] is calculated
as:

𝒖𝒊 ∣𝒌 = 𝒙𝒊 ∣𝒌 + 𝐹𝑖 (𝒗 𝒊 ∣𝒌 − 𝒙𝒊 ∣𝒌 ) (5)

Fig. 1: Suggested use from literature for Cr values. With

It is recommended 𝐹𝑖 = 0.5⋅(𝐹 +1) as a good initial choice
[13].
3) Selection: A greedy selection is used, where a child
substitutes its father in the next generation if it has a better or
equal fitness function value.
Cr=1.0, algorithm is fully rotationally invariant.
In this way, the algorithm will be able to detect if high
values of 𝐶𝑟 are useful, and furthermore, if a rotationally

1552
invariant crossover is required. A minimum base for Λ𝐶𝑟
around its median value is incorporated to avoid stagnation
around a single value. Also, a threshold 𝛼𝐶𝑟 is defined to use
directly the line recombination crossover if 𝑓𝐶𝑟 (𝒙)∣𝑘 > 𝛼𝐶𝑟
3) Population management: An initial population 𝑃 (0)
with 𝑁𝑝 individuals is randomly selected from the decision
space. A subpopulation of 𝑁𝑠 individuals is used in each
generation step. A population refreshment is incorporated
when the difference of the interquartil range (𝑰𝑸𝑹) is below
∗
−𝐿𝑜𝑤𝑒𝑟 ∗
𝑽ˆ = 𝑈 𝑝𝑝𝑒𝑟𝛾𝑣𝑎𝑟 where 𝑈 𝑝𝑝𝑒𝑟∗ and 𝐿𝑜𝑤𝑒𝑟∗ are the
upper and lower bounds since the last population refreshment.
The refreshed population is initialized with 𝑁𝑅 vector solu-
tions in the range 𝒙𝑚𝑒𝑑𝑖𝑎𝑛 ± 𝛾𝑣𝑎𝑟 ⋅ 𝑽ˆ ; the new initialization
range for the next refreshment (if required) will be defined by
𝒙𝑚𝑒𝑑𝑖𝑎𝑛 ±𝛾𝑣𝑎𝑟 ⋅ 𝑽ˆ bounded to a minimum 𝑽ˆ𝑚𝑖𝑛 . Interquartile
difference is used instead variance, due to its robustness
properties to exclude outliers.
The main steps of the hybrid DE are:
S1: Initialize 𝑃 (0) with 𝑁𝑝 individuals randomly se-
lected from the searching space.
S2: Evaluate initial population 𝑃 (0)
S3: For 𝑘 = 1 : MaxGen or convergence criterion
reached
S3.1: Select randomly a subpopulation of 𝑁𝑆 (𝑘)
individuals with proposed solutions on
𝑃 (𝑘).
S3.2: Apply the fundamental DE operators on
subpopulation 𝑁𝑆 (𝑘) to get the offspring
𝑂(𝑘):
1) Perform mutation operator (equation 3)
with F selected from the triangular dis-
tribution.
2) Perform crossover operator (equation 4)
with 𝐶𝑟 selected from triangular dis-
tribution. If 𝐶𝑟 > 𝛼𝐶𝑟 , the arithmetic
recombination is used.
S3.4: For every child, perform a local search with
a probability of 𝛼𝐿𝑆
S3.3: Evaluate offspring 𝑂(𝑘); if child ≤ parent,
the parent is substituted by its child.
S3.4: Record the number of success. If the cu-
mulative number of success is bigger than
𝛾𝑠𝑢𝑐𝑐𝑒𝑠𝑠 , reset the cumulative number and
recalculate the triangular distribution for Cr.
S3.5: If 𝑰𝑸𝑹 < 𝑽ˆ , refresh the population with
𝑁𝑅 individuals from 𝒙𝑚𝑒𝑑𝑖𝑎𝑛 ± 𝛾𝑣𝑎𝑟 ⋅ 𝑉ˆ .
S4: Algorithm terminates.
III. E XPERIMENTAL S ETUP
The algorithm was implemented in Matlab⃝ c with an stan-
dard computer with 1.6GHz processor and 1.9 GB RAM. The
following tuning parameters are selected for the 𝐷𝐸 − Λ𝐶𝑟
proposal:
Local Search parameter:
1
∙ 𝛼𝐿𝑆 = 100⋅𝑛 .
1553
Adaptive mechanism:
∙ 𝛾𝑠𝑢𝑐𝑐𝑒𝑠𝑠 = 15;
∙ 𝛼𝐶𝑟 = 0.95;
∙ Minimum base for Λ𝐶𝑟 : 𝑚𝑒𝑑𝑖𝑎𝑛 ± 0.1.
∙ Λ𝐹 = {0.3, 0.4, 0.5}.
∙ 𝐹𝑖 = 0.5 ⋅ (1 + 0.5) = 0.75
Population Management:
∙ 𝑁𝑅 = 2 ⋅ 𝑛 (bounded to [20, 120]).
∙ 𝑁𝑝 = 5 ⋅ 𝑛 (bounded to [50, 300]).
∙ 𝑁𝑠 = 5 ⋅ 𝑛 (bounded to [50, 300]).
∙ 𝛾𝑣𝑎𝑟 = 3.
∙ 𝑽ˆ𝑚𝑖𝑛 = 0.05 ⋅ 𝒙𝑢𝑝𝑝𝑒𝑟 − 𝒙𝑙𝑜𝑤𝑒𝑟 .
IV. D ISCUSSION
In table II, the best result obtained by 𝐷𝐸 − Λ𝐶𝑟 proposal
for each problem is shown, whilst in tables III, IV, V,
VI and VII the obtained results for FES=5e4, FES=10e4,
FES=15e4 are presented. The performance of some problems
could be improved with different parameter tuning values. The
competition discourages to use a different set of parameters for
each problems and therefore, general values are used instead.
From reported results in the literature it is possible to notice
some remarks on the algorithm performance:
∙ Regarding problem P01, the 𝐷𝐸 − Λ𝐶𝑟 proposal is
capable to find low values for parameter estimation for
FM sound waves, and it is competitive with other reported
values [17].
∙ For problem subset P11, the 𝐷𝐸 − Λ𝐶𝑟 algorithm
structure is able to detect the strong dependency on
search variables. This strong dependency is due to the
interdependency among variables, since the feasible space
demands to have an specific load. It is not possible
to find a better value than the reported in other works
[18], but it is possible to have competitive results. As
mentioned earlier, the constraint handling is incorporated
in the problem definitions. A different constraint handling
mechanism with the proposed algorithm could lead to
better results.
∙ Regarding European Space Agency (ESA) problems
(problems P12 and P13), notice that the algorithm has
problems to find a reasonable solution for messenger
problem (P12) with the allowed FEs. The better value
reported from the ESA is 2.970 km/sec (recently found).
For Cassini problem (P13), some executions are able to
reach values below 9.0. The best reported results is 8.383
kg km/sec by MIDACO algorithm [19], which required
50 days to reach its solution.
V. C ONCLUSION
The results of the proposed 𝐷𝐸 − Λ𝐶𝑟 algorithm for the
CEC-2011 special session on real world optimization are
presented. The problem set consist on 13 problems with
diverse topics, including control systems, spacecraft trajectory
optimization, economic load dispatch, engineering design,
among others.
 TABLE II: Best results achieved. [10] B. Qu and P. Suganthan, “Multi-objective evolutionary algorithms based
Problem Best value (this paper) on the summation of normalized objectives and diversified selection,”
Information Sciences, vol. 180, no. 17, pp. 3170 – 3181, 2010, including
P01 7.2093E-15
Special Section on Virtual Agent and Organization Modeling: Theory
P02 -2.8423E+01
and Applications.
P03 1.1515E-05
[11] Q. Zhang and H. Li, “Moea/d: A multiobjective evolutionary algorithm
P04 1.3772E+01 based on decomposition,” Evolutionary Computation, IEEE Transactions
P05 -3.6845E+01 on, vol. 11, no. 6, pp. 712 – 731, 2007.
P06 -3.6845E+01 [12] G. Reynoso-Meza, “Design, coding and implementation of a
P07 6.6591E-01 multiobjective optimization algorithm based on differential evolution
P08 2.2000E+02 with spherical pruning: applications for system identification and
P09 3.2527E+05 controller tuning.” Master’s thesis, Universidad Politécnica de Valencia.,
P10 -2.1601E+01 2009. [Online]. Available: http://personales.alumno.upv.es/gilreyme
P11.1 4.4847E+04 [13] K. V. Price, An introduction to differential evolution. Maidenhead, UK,
P11.2 1.0354E+06 England: McGraw-Hill Ltd., UK, 1999, pp. 79 – 108.
P11.3 1.5445E+04 [14] N. Noman and H. Iba, “Accelerating differential evolution using an
P11.4 1.8028E+04 adaptive local search,” Evolutionary Computation, IEEE Transactions
P11.5 3.2730E+04 on, vol. 12, no. 1, pp. 107 –125, feb. 2008.
P11.6 1.2313E+05 [15] F. Neri and V. Tirronen, “Scale factor local search in differential evolu-
P11.7 1.7124E+06 tion,” Memetic Computing, vol. 1, pp. 153–171, 2009, 10.1007/s12293-
P11.8 9.2280E+05 009-0008-9.
P11.9 9.2717E+05 [16] A. LaTorre, S. Muelas, and J.-M. Peña, “A mos-based dynamic memetic
P11.10 9.2303E+05 differential evolution algorithm for continuous optimization: a scalability
P12 1.1814E+01 test,” Soft Computing - A Fusion of Foundations, Methodologies and
P13 8.9624E+00 Applications, pp. 1–13, 2010, 10.1007/s00500-010-0646-3.
[17] F. Herrera and M. Lozano, “Gradual distributed real-coded genetic
algorithms,” Evolutionary Computation, IEEE Transactions on, vol. 4,
no. 1, pp. 43 –63, Apr. 2000.
[18] L. Coelho and V. Mariani, “Combining of chaotic differential evolution
The 𝐷𝐸 − Λ𝐶𝑟 proposal use a local search routine to and quadratic programming for economic dispatch optimization with
improve converge and an adaptive parameter mechanism for valve-point effect,” Power Systems, IEEE Transactions on, vol. 21, no. 2,
pp. 989 – 996, May 2006.
the crossover operator. According with the reported results, the [19] M. Schlueter, J.-J. Ruckmann, and M. Gerdts, “Non-linear mixed-
algorithm is capable to find competitive solutions compared integer-based optimisation technique for space applications,” Poster ESA
with the reported results from the literature. 1st Networking/Partnering Day 2010, January 2010.

ACKNOWLEDGMENT
This work was partially supported by the FPI-2010/19 grant
from the Universitat Politècnica de València and the project
DPI2008-02133/DPI from the Spanish Ministry of Science and
Innovation.
R EFERENCES
[1] S. Das and P. Suganthan, “Problem definitions and evaluation criteria for
cec 2011 competition on testing evolutionary algorithms on real world
optimization problems,” Jadavpur university and Nanyang Technological
University, Tech. Rep., 2011.
[2] R. Storn and K. Price, “Differential evolution: A simple and efficient
heuristic for global optimization over continuous spaces,” Journal of
Global Optimization, vol. 11, pp. 341 – 359, 1997.
[3] R. Storn, “Sci: Differential evolution research: Trends and open ques-
tions,” U. K. C. (Ed.), Ed. Springer, Heidelberg, 2008, vol. LNCS 143,
pp. 1 – 31.
[4] S. Das and P. N. Suganthan, “Differential evolution: A survey of
the state-of-the-art,” Evolutionary Computation, IEEE Transactions on,
2010.
[5] S. Kukkonen and J. Lampinen., “Gde3: The third evolution step on gen-
eralized differential evolution,” vol. 1. IEEE Congress on Evolutionary
Computation (CEC 2005), 2005, pp. 443 – 450.
[6] J. Brest, B. Boskovic, S. Greiner, V. Zumer, and M. S. Maucec, “Per-
formance comparison of self-adaptive and adaptive differential evolution
algorithms,” vol. 11, no. 7, pp. 617 – 629, 2007.
[7] A. Qin and P. Suganthan, “Self-adaptive differential evolution algorithm
for numerical optimization,” vol. 2. Evolutionary Computation, 2005.
The 2005 IEEE Congress on, 2005, pp. 1785 – 1791.
[8] V. Huang, A. Qin, P. Suganthan, and M. Tasgetiren, “Multi-objective
optimization based on self-adaptive differential evolution algorithm,” pp.
324 – 331, 2007.
[9] R. Mallipeddi, P. Suganthan, Q. Pan, and M. Tasgetiren, “Differential
evolution algorithm with ensemble of parameters and mutation strate-
gies,” Applied Soft Computing, vol. 11, no. 2, pp. 1679 – 1696, 2011,
the Impact of Soft Computing for the Progress of Artificial Intelligence.

1554
TABLE III: Function values achieved when FES=5e4, FES=10e4, FES=15e4 for parameter estimation for FM sound waves
(P01), Lennard-Jones Potential (P02), bifunctional catalyst blend optimal control (P03), optimal control of a non-linear stirred
tank reactor (P04) and Tersoff potential minimization Si(B) problems.
P01 P02 P03 P04 P05
Worst 1.4813E+01 -2.2032E+01 1.1515E-05 2.1025E+01 -2.4250E+01
Median 3.8779E-09 -2.7447E+01 1.1515E-05 2.0959E+01 -3.2957E+01
5e4 FES Best 1.6147E-10 -2.8423E+01 1.1515E-05 1.3772E+01 -3.6845E+01
Mean 2.1870E+00 -2.6525E+01 1.1515E-05 1.9524E+01 -3.2358E+01
Std 4.6326E+00 1.7450E+00 2.2930E-17 2.4861E+00 3.2663E+00
Worst 1.1757E+01 -2.6362E+01 1.1515E-05 2.1012E+01 -3.1484E+01
Median 5.0886E-11 -2.7545E+01 1.1515E-05 2.0958E+01 -3.4252E+01
10e4 FES Best 3.5443E-12 -2.8423E+01 1.1515E-05 1.3772E+01 -3.6845E+01
Mean 1.0978E+00 -2.7527E+01 1.1515E-05 1.9343E+01 -3.4237E+01
Std 3.1751E+00 6.5994E-01 1.1207E-17 2.6044E+00 1.5187E+00
Worst 1.1757E+01 -2.6443E+01 1.1515E-05 2.1002E+01 -3.1484E+01
Median 1.2362E-11 -2.7545E+01 1.1515E-05 1.7487E+01 -3.4870E+01
15e4 FES Best 7.2093E-15 -2.8423E+01 1.1515E-05 1.3772E+01 -3.6845E+01
Mean 8.7697E-01 -2.7731E+01 1.1515E-05 1.7339E+01 -3.4720E+01
Std 3.0439E+00 4.9035E-01 6.1087E-18 2.9761E+00 1.4469E+00

TABLE IV: Function values achieved when FES=5e4, FES=10e4, FES=15e4 for Tersoff potential minimization Si(C) (P06)
spread spectrum radar (C07), transmission network expansion planning (C08), large scale transmission pricing problem (C09)
and circular antenna array design (C10) problems.
P06 P07 P08 P09 P10
Worst -2.4495E+01 1.1839E+00 2.2000E+02 1.3610E+06 -1.0889E+01
Median -3.2897E+01 9.8081E-01 2.2000E+02 3.4789E+05 -1.6165E+01
5e4 FES Best -3.6845E+01 8.3380E-01 2.2000E+02 3.2528E+05 -2.1572E+01
Mean -3.2239E+01 9.8858E-01 2.2000E+02 4.5612E+05 -1.6736E+01
Std 3.5197E+00 1.0315E-01 0.0000E+00 2.6085E+05 4.0332E+00
Worst -3.1693E+01 1.1263E+00 2.2000E+02 5.6363E+05 -1.0939E+01
Median -3.4165E+01 9.3720E-01 2.2000E+02 3.3088E+05 -1.6198E+01
10e4 FES Best -3.6845E+01 7.1698E-01 2.2000E+02 3.2527E+05 -2.1586E+01
Mean -3.4183E+01 9.2967E-01 2.2000E+02 3.4660E+05 -1.6751E+01
Std 1.3086E+00 1.0051E-01 0.0000E+00 4.7757E+04 4.0388E+00
Worst -3.4165E+01 1.0361E+00 2.2000E+02 3.6130E+05 -1.0940E+01
Median -3.4316E+01 8.9739E-01 2.2000E+02 3.3088E+05 -1.6198E+01
15e4 FES Best -3.6845E+01 6.6591E-01 2.2000E+02 3.2527E+05 -2.1601E+01
Mean -3.5033E+01 8.8477E-01 2.2000E+02 3.3547E+05 -1.6756E+01
Std 1.0287E+00 1.0571E-01 0.0000E+00 1.1809E+04 4.0437E+00

TABLE V: Function values achieved when FES=5e4, FES=10e4, FES=15e4 for Economic dispatch load problems (P11.1 ∽
P11.5 ).
P11.1 P11.2 P11.3 P11.4 P11.5
Worst 5.1872E+04 1.9581E+06 1.5448E+04 1.8196E+04 3.2953E+04
Median 4.7680E+04 1.0407E+06 1.5446E+04 1.8088E+04 3.2842E+04
5e4 FES Best 4.5458E+04 1.0376E+06 1.5445E+04 1.8028E+04 3.2773E+04
Mean 4.7747E+04 1.1586E+06 1.5446E+04 1.8099E+04 3.2850E+04
Std 1.3962E+03 2.9044E+05 7.9715E-01 4.0566E+01 5.0620E+01
Worst 4.8019E+04 1.2968E+06 1.5448E+04 1.8196E+04 3.2939E+04
Median 4.6316E+04 1.0376E+06 1.5446E+04 1.8088E+04 3.2810E+04
10e4 FES Best 4.4847E+04 1.0354E+06 1.5445E+04 1.8028E+04 3.2762E+04
Mean 4.6423E+04 1.0498E+06 1.5446E+04 1.8098E+04 3.2816E+04
Std 7.9974E+02 5.1927E+04 8.5993E-01 4.0541E+01 3.9643E+01
Worst 4.8019E+04 1.2810E+06 1.5448E+04 1.8196E+04 3.2854E+04
Median 4.6232E+04 1.0375E+06 1.5446E+04 1.8087E+04 3.2789E+04
15e4 FES Best 4.4847E+04 1.0354E+06 1.5445E+04 1.8028E+04 3.2730E+04
Mean 4.6321E+04 1.0490E+06 1.5446E+04 1.8096E+04 3.2790E+04
Std 8.5566E+02 4.8789E+04 8.4385E-01 3.9913E+01 3.0447E+01

1555
TABLE VI: Function values achieved when FES=5e4, FES=10e4, FES=15e4 for Economic dispatch load problems (P11.6 ∽
P11.10 ).
P11.6 P11.7 P11.8 P11.9 P11.10
Worst 1.2553E+05 1.9712E+06 9.2842E+05 9.4618E+05 9.5266E+05
Median 1.2445E+05 1.8502E+06 9.2519E+05 9.3229E+05 9.2550E+05
5e4 FES Best 1.2359E+05 1.7124E+06 9.2390E+05 9.2717E+05 9.2312E+05
Mean 1.2440E+05 1.8653E+06 9.2533E+05 9.3282E+05 9.2710E+05
Std 5.0530E+02 5.7404E+04 1.1126E+03 3.9035E+03 6.7496E+03
Worst 1.2520E+05 1.9250E+06 9.2680E+05 9.3507E+05 9.2597E+05
Median 1.2404E+05 1.8483E+06 9.2413E+05 9.3079E+05 9.2427E+05
10e4 FES Best 1.2313E+05 1.7124E+06 9.2307E+05 9.2717E+05 9.2312E+05
Mean 1.2400E+05 1.8391E+06 9.2442E+05 9.3061E+05 9.2436E+05
Std 4.4568E+02 4.5459E+04 8.2878E+02 1.7486E+03 8.0871E+02
Worst 1.2520E+05 1.9250E+06 9.2542E+05 9.3348E+05 9.2481E+05
Median 1.2390E+05 1.8483E+06 9.2389E+05 9.2999E+05 9.2376E+05
15e4 FES Best 1.2313E+05 1.7124E+06 9.2280E+05 9.2717E+05 9.2303E+05
Mean 1.2389E+05 1.8386E+06 9.2393E+05 9.3001E+05 9.2382E+05
Std 4.5430E+02 4.5074E+04 7.4822E+02 1.6287E+03 5.1463E+02

TABLE VII: Function values achieved when FES=5e4, FES=10e4, FES=15e4 for Messenger (P12) and Cassini (P13) problems.
P12 P13
Worst 2.1172E+01 3.0450E+01
Median 1.7047E+01 2.1860E+01
5e4 FES Best 1.1814E+01 1.0544E+01
Mean 1.6943E+01 2.0483E+01
Std 1.9908E+00 4.4830E+00
Worst 1.8486E+01 2.3129E+01
Median 1.5733E+01 1.5778E+01
10e4 FES Best 1.1814E+01 8.9636E+00
Mean 1.5888E+01 1.6422E+01
Std 1.4035E+00 3.8442E+00
Worst 1.7891E+01 2.1052E+01
Median 1.5219E+01 1.4412E+01
15e4 FES Best 1.1814E+01 8.9624E+00
Mean 1.5360E+01 1.4909E+01
Std 1.2133E+00 2.7634E+00

1556